CN105427341A - Multi-variation level set-based multi-target detection method for complicated background video images - Google Patents

Abstract

The invention discloses a multi-variation level set-based multi-target detection method for complicated background video images. The method comprises the following steps: (1) inputting video images in an image format; (2) selecting three image frames; (3) obtaining a binary image; (4) removing the interference noise; (5) obtaining a zero level set of a moving object; (6) updating the zero level set of the moving object by using a multi-variation level set method; (7) judging whether the updated zero level set of the moving object coincides with the edge of a moving object in the second frame of image; and (8) outputting the detection result of the moving object in the second frame of image. According to the method provided by the invention, the problem that the zero level sets of the moving objects are fixed curves is solved, and the deficiency that the detection results disturb one another as image region global information is adopted during the detection of a plurality of moving objects is solved; and the method has relatively high detection correctness and can be applied to the detection of a plurality of moving objects in video images.

Description

Multi-variation level set-based complex background video image multi-target detection method

Technical Field

The invention belongs to the technical field of image processing, and further relates to a multi-target detection method for a complex background video image based on a multi-variation level set in the technical field of multi-target detection of video images. The method can be used for identifying whether the moving target exists in the video image and detecting a plurality of identified moving targets under the complex image background.

Background

The moving object detection is to detect a change area from a sequence image, extract a moving object from a background image, and divide a foreground area changed in the sequence image from each frame image. In general, post-processing procedures such as object classification, tracking and behavior understanding only consider pixel regions corresponding to moving objects in an image, and therefore correct detection and segmentation of moving objects are very important for post-processing. However, due to the dynamic changes of the scene, such as the influence of weather, light, shadow and cluttered background interference, the detection and segmentation of the moving object are difficult. Motion detection is divided into static and dynamic backgrounds depending on whether the camera is held still. Most of video monitoring systems have fixed cameras, so moving object detection algorithms in static backgrounds are widely concerned, and commonly used methods include a frame difference method, an optical flow method, a background difference method and the like.

Yangli, Zunongguo, Hujing, Liu et al propose a moving target level set extraction method based on a c-v model in a paper 'moving target level set extraction method based on a c-v model' (Taiyuan science and technology university newspaper, vol. 33, No. 4 of 2012, 8). The method comprises the steps of initially detecting a motion area by using an improved interframe difference method, judging a motion target pixel in a current video by using an adaptive threshold value through subtraction of adjacent video frames, removing noise interference through morphological processing, and constructing a level set c-v model extracted from a motion target contour by defining a minimum energy function, thereby realizing detection of the motion target. The method has the disadvantages that the zero level set of the moving target is performed by adopting a fixed curve, and the method can cause the non-moving target to be falsely detected as the moving target, thereby reducing the detection accuracy.

Song Li Ye, Frielin et al in their published paper "multiphase horizontal set image segmentation method without reinitializing the level set" ("Shandong university Collection, 3/2012, Vol. 27, No. 1) propose a multiphase horizontal set image segmentation method without reinitializing the level set. According to the regional competition thought of the multi-phase model proposed by Vese-Chan, the method combines the energy function of the image regional global information as the external energy item of the model, introduces the internal deformation energy constraint level set function to approximate the symbolic distance function, avoids the process of reinitializing the level set function, improves the calculation speed and the segmentation effect, and simultaneously realizes the detection of the multi-phase target in the image. The method has the disadvantages that the method can not detect the local area by combining the global information of the image area, and when the method is used for detecting the zero level of a plurality of moving targets, each moving target zero level set is interfered by other moving target zero level sets to influence the detection accuracy.

Liu Lixiong, Chen Meng Juan et al proposed a multi-target image segmentation method based on level sets in the patent of "a multi-target image segmentation method based on level sets" (application patent No. 201310029907.6, publication No. CN 103093473A). According to the method, one or more closed curves are drawn on an image to be segmented as an initial contour, and then an active contour model based on a region is used for carrying out iterative evolution on the initial contour, so that a contour curve of a target is obtained finally. The active contour model based on the region fully considers the local gray information of the image, so that the image with uneven gray can be segmented; the model is popularized to multiple stages, and multi-target images can be segmented. The method has the disadvantages that the positions of the zero level sets are randomly selected, the number of the zero level sets is not determined according to the number of the targets, and the background is judged to be the target for multi-target detection in a complex scene, so that the detection accuracy is influenced.

Disclosure of Invention

The invention aims to provide a multi-target detection method for a complex background video image based on a multi-variation level set aiming at the defects of the prior art. The invention reduces the detection complexity, better avoids the situation that the background is mistakenly detected as the moving target under the complex background, and improves the detection accuracy.

The method comprises the following specific steps:

(1) inputting video images in a video format:

(2) selecting continuous three-frame video image with obvious moving object from input video image, defining it as first frame video image I₁Second frame video image I₂Third frame video image I₃；

(3) Obtaining a binary image:

performing interframe difference on the selected continuous three-frame image by adopting an interframe difference method to obtain a binary image B;

(4) removing interference noise:

denoising other interference noises except the target in the binary image by adopting a morphological function to obtain a denoised binary image I;

(5) obtaining a zero level set of the moving object:

(5a) clustering the binary image I after the noise is removed by adopting a K-means clustering algorithm to respectively obtain the sum of the distances from all data points in a clustering class to the clustering center, the number of all data points in the clustering class and the clustering center;

(5b) taking the result of comparing the sum of the distances from all the data points in the cluster category to the cluster center with the number of all the data points in the cluster category as the radius, and taking the cluster center point as the center of a circle to obtain a circle which is the zero level set phi of the moving target_k；

(6) Updating a moving target zero level set by using a multi-variation level set method:

(6a) the internal energy of the zero level set of the moving object is calculated according to the following formula:

$W\left(\mathrm{\φ}\right)=\frac{\mathrm{\μ}}{2}\underset{k=1}{\overset{K}{\Σ}}{\∫}_{\mathrm{\Ω}}{\left(\right|\▿{\mathrm{\φ}}_k|-1)}^{2}dxdy$

wherein W (phi) represents the internal energy of the zero level set of the moving object, and phi represents the zero level set of the moving object_kMu represents the constraint coefficient of the internal energy, the value range of the constraint coefficient is more than 0 and less than 0.25, and k represents the constraint coefficient of the noise removalThe number of cluster categories obtained by clustering data points in the post-vocal binary image I, K being 1,2, …, K representing the total number of moving objects, ∑ representing summation operation, ^ integral representing integral operation, and Ω representing the second frame video image I₂The plane set, |, represents the absolute value operation, ▽ represents the gradient operator operation, phi_kZero level set representing moving object, x represents zero level set phi of moving object_kThe coordinate value of the horizontal axis of (a) y represents a zero level set phi of the moving object_kThe coordinate value of the longitudinal axis of (c);

(6b) obtaining a second frame video image I according to the following formula₂Edge of middle moving object:

$g=\frac{1}{1+|\▿G*I_2{|}^2}$

wherein g represents the second frame video image I₂The edge of the medium moving target, |, represents the operation of taking absolute value, ▽ represents the operation of solving gradient operator, G represents Gaussian kernel, I represents convolution operation₂Representing a second frame of video image of the three consecutive frames of video images;

(6c) the length of the zero level set of the moving object is calculated according to the following formula:

$L_g\left(\mathrm{\φ}\right)=\underset{k=1}{\overset{K}{\Σ}}{\∫}_{\mathrm{\Ω}}g\mathrm{\δ}\left({\mathrm{\φ}}_k\right)|\▿{\mathrm{\φ}}_k|dxdy$

wherein L is_g(phi) denotes the length of the zero level set of the moving object, and g denotes the second frame video image I₂The edge of the middle moving object, phi, represents the zero level set phi of the moving object_kSet of (a) phi_kRepresents a zero level set of the moving object, K represents a serial number of a cluster category obtained by clustering data points in the binary image I after the noise is removed, K ═ 1,2, …, K represents the total number of the moving object, ∑ represents a summation operation, ^ represents an integral operation, and Ω represents a second frame video image I₂The plane set of (1) represents a Dirac function of a single variable, | · | represents an absolute value taking operation, ▽ represents a gradient operator solving operation, and x represents a zero level set phi of a moving target_kThe coordinate value of the horizontal axis of (a) y represents a zero level set phi of the moving object_kThe coordinate value of the longitudinal axis of (c);

(6d) the area of the moving target region is calculated as follows:

$A_g\left(\mathrm{\φ}\right)=\underset{k=1}{\overset{K}{\Σ}}{\∫}_{\mathrm{\Ω}}gH(-{\mathrm{\φ}}_k)dxdy$

wherein A is_g(phi) denotes the area of the moving object region, and g denotes the second frame video image I₂The edge of the middle moving object, phi, represents the zero level set phi of the moving object_kSet of (a) phi_kRepresents a zero level set of the moving object, K represents a serial number of a cluster category obtained by clustering data points in the binary image I after the noise is removed, K ═ 1,2, …, K represents the total number of the moving object, ∑ represents a summation operation, ^ represents an integral operation, and Ω represents a second frame video image I₂H represents the Heaviside function, x represents the zero level set phi of the moving object_kThe coordinate value of the horizontal axis of (a) y represents a zero level set phi of the moving object_kThe coordinate value of the longitudinal axis of (c);

(6e) the external energy of the zero level set of the moving object is calculated according to the following formula:

${\mathrm{\ϵ}}_{g,{\mathrm{\λ}}_k,v_k}\left(\mathrm{\φ}\right)=\underset{k=1}{\overset{K}{\Σ}}{\mathrm{\λ}}_k{L}_g\left(\mathrm{\φ}\right)+\underset{k=1}{\overset{K}{\Σ}}{v}_k{A}_g\left(\mathrm{\φ}\right)$

wherein,representing the external energy of the zero level set of the moving object, g representing the second frame video image I₂Edge of medium moving object, λ_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer less than 10, K denotes a serial number of a cluster type obtained by clustering data points in the binary image I after the noise is removed, K is 1,2, …, K denotes a total number of moving objects, v denotes a number of moving objects, and K denotes a number of clusters_kA constraint coefficient for representing the area of the moving target region, phi represents the zero level set phi of the moving target_kA value of-5 < v_k< 5 and v_kNot equal to 0, ∑ denotes a summation operation, L_g(phi) represents the length of the zero level set of the moving object, A_g(phi) represents the area of the moving target region;

(6f) the total energy of the zero level set of the moving object is calculated according to the following formula:

$\mathrm{\&epsiv;}\left(\mathrm{\&phi;}\right)=W\left(\mathrm{\&phi;}\right)+{\mathrm{\&epsiv;}}_{g,{\mathrm{\&lambda;}}_k,v_k}\left(\mathrm{\&phi;}\right)$

wherein (phi) represents the total energy of the zero level set of the moving target, and phi represents the zero level set phi of the moving target_kW (phi) denotes the zero level set of moving objectsThe energy of the interior is such that,representing the external energy of the zero level set of the moving object, g representing the second frame video image I₂Contour of medium moving object, λ_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer less than 10, k represents the serial number of the cluster category obtained after clustering the data points in the binary image I after removing the noise, v_kA constraint coefficient representing the area of the moving target region, and the value of the constraint coefficient is-5 < v_k< 5 and v_k≠0；

(6g) Carrying out level set evolution on the moving target zero level set according to the following formula to obtain an updated moving target zero level set:

$\frac{\&part;{\mathrm{\&phi;}}_k}{\&part;t}=\mathrm{\&mu;}\&lsqb;{\mathrm{\&Delta;\&phi;}}_k-div\left(\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}\right)\&rsqb;+\mathrm{\&delta;}\left({\mathrm{\&phi;}}_k\right)\&lsqb;\underset{k=1}{\overset{K}{\&Sigma;}}{\mathrm{\&lambda;}}_{k}div\left(g\frac{\&dtri;{\mathrm{\&phi;}}_k}{|\&dtri;{\mathrm{\&phi;}}_k|}\right)\&rsqb;+\underset{k=1}{\overset{K}{\&Sigma;}}{v}_{k}g\mathrm{\&delta;}\left({\mathrm{\&phi;}}_k\right)$

wherein,represents the updated zero level set of the moving object,denotes a derivative operation phi_kRepresents the zero level set of the moving object, phi represents the zero level set of the moving object_kK represents the serial number of a cluster class obtained by clustering data points in the binary image I after noise removal, K is 1,2, …, K represents the total number of moving targets, t represents an iteration step length, mu represents a constraint coefficient for internal energy, the value range of the constraint coefficient is 0 & ltmu & lt 0.25, delta represents a Laplacian operator operation, div (·) represents the divergence of a solved vector, ▽ represents a gradient operator operation, | · | represents an absolute value operation, represents a Dirac function of a single variable, ∑ represents a summation operation, and λ_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer < 10, g denotes a second frame video image I₂Edge of medium moving object, v_kA constraint coefficient representing the area of the target region, the value of which is-5 < v_k< 5 and v_k≠0；

(7) Judging whether the updated moving object zero level set is corresponding to the second frame video image I₂If the edges of the middle moving target coincide, executing the step (8), otherwise, executing the step (6);

(8) outputting a second frame video image I₂And (5) detecting the moving target.

Compared with the prior art, the invention has the following advantages:

firstly, the invention adopts the k-means clustering algorithm to obtain the zero level set of the moving target and the position of the zero level set of the moving target, and determines the number of the zero level sets of the moving target according to the number of the moving target, thereby avoiding that the zero level set of the moving target in the prior art is obtained by a fixed curve, which causes that a non-moving target is falsely detected as the moving target, and the position of the zero level set in the prior art is randomly selected, the number of the zero level sets is not determined according to the number of the moving target, for multi-target detection in a complex scene, the background can be judged as the target, and the detection capability is insufficient, thus improving the accuracy of detecting the moving target in a video image.

Secondly, because the invention adopts the multi-variation level set method to update the moving target zero level set, the defect that the prior art adopts the global information of the image area and can not detect the local area is avoided, when a plurality of moving targets are detected, each moving target zero level set can be interfered by other moving target zero level sets, the position and the number of the moving target zero level sets are comprehensively considered, and the accuracy of the multi-target detection of the complex background in the video image is improved.

Drawings

FIG. 1 is a flow chart of the present invention;

fig. 2 is a diagram illustrating the effect of detecting a moving object in a video image containing a plurality of moving objects according to the present invention.

Detailed Description

The present invention is described in further detail below with reference to the attached drawing figures.

The steps of the present invention will be described in further detail with reference to fig. 1.

Step 1: a video image in a video format is input.

Step 2: selecting continuous three-frame video image with obvious moving object from input video image, defining it as first frame video image I₁Second frame video image I₂Third frame video image I₃。

And step 3: and obtaining a binary image.

And performing interframe difference on the selected continuous three-frame image by adopting an interframe difference method to obtain a binary image B.

The specific steps of the interframe difference method are as follows:

using the first frame video image I of the selected continuous three frame video images₁Subtracting the second frame video image I₂Obtaining a differential image B¹。

Using the second frame video image I of the selected continuous three frame video images₂Subtracting the third frame video image I₃Obtaining a differential image B²。

Two differentiated images B¹And B²And (5) carrying out phase AND to obtain a binary image B.

In the present invention, a binary image obtained by performing inter-frame difference operation on the selected continuous three-frame video images by using the above inter-frame difference algorithm is shown in fig. 2 (a). In fig. 2(a), the moving object is a white region, and the background is a black region. As can be seen from fig. 2, there is noise interference in the video image, which mistargets the background, and therefore, a small number of white dots are included in the black background region.

And 4, step 4: and removing interference noise.

And denoising other interference noises except the target in the binary image by adopting a morphological function to obtain a denoised binary image I.

In the invention, the obtained binary image is processed by a bewareaopen function with a morphological function of MATLAB software, and the obtained binary image after noise removal is shown as an attached figure 2 (b). In fig. 2(b), the moving object is a white region, and the background is a black region. As can be seen from fig. 2(b), white dots existing in the background area in the binary image are removed, and noise interference existing in the background in the binary image is removed, so that the influence of background noise on target detection is reduced, and the detection accuracy is improved.

And 5: a zero level set of moving objects is obtained.

And clustering the binary image I after the noise is removed by adopting a K-means clustering algorithm to respectively obtain the sum of the distances from all data points in the clustering class to the clustering center, the number of all data points in the clustering class and the clustering center.

The k-means clustering algorithm comprises the following specific steps:

and randomly selecting K central points from the binary image I after the noise is removed, wherein K represents the total number of the moving targets.

Traversing all data points in the binary image I after the noise is removed according to the following formula, and dividing each data point into the nearest central point:

$\begin{array}{cc}\min & J\left(C\right)=\underset{k=1}{\overset{K}{\&Sigma;}}\underset{x_i\&Element;C_k}{\&Sigma;}{m}_{ki}dist{(x_i,s_k)}^2\end{array}$

j (C) represents the distance from each data point in the binary image I after the noise is removed to the center point of the division, C represents a cluster category set obtained after the data points in the binary image I after the noise is removed are divided, min (·) represents minimum value taking operation, ∑ represents summation operation, m (·) represents summation operation, m represents the sum of the data points in the binary image I after the noise is removed, and the sum of the data points in the binary image I is equal to m_kiRepresenting coefficients, I representing the sequence numbers of data points in the denoised binary image I, k representing the denoised binary image IThe serial number of the cluster type obtained by dividing the data points in the binary image I, K is 1,2, …, K represents the total number of moving targets, dist (·) represents a distance function, and x represents_iRepresenting the ith data point, s, in the denoised binary image I_kA cluster center point C representing the kth class obtained by dividing the binary image I after removing the noise_kThe class k category obtained by dividing the binary image I from which the noise has been removed is shown, and ∈ shows the class belongs to the operation.

Calculating the average value of each cluster category obtained after dividing the binary image I after removing the noise according to the following formula, and taking the average value as a new cluster center point:

$s_k=\frac{1}{n_k}\underset{i=1}{\overset{n_k}{\&Sigma;}}{x}_i^k$

wherein s is_kRepresenting the clustering center point of the kth class obtained after the division of the binary image I after the noise removal, k representing the serial number of the class obtained after the division of the data points in the binary image I after the noise removal, n_kIndicating the number of data points in the kth cluster class obtained by dividing the data points in the binary image I after the noise is removed, ∑ indicating the summation operation,and the data points in the binary image I after the noise is removed are divided to obtain the ith data point in the kth clustering class, and I represents the serial number of the data point in the binary image I after the noise is removed.

To gatherThe result of the comparison of the sum of the distances from all the data points in the class type to the cluster center of the class type and the number of all the data points in the cluster type is used as a radius, and a circle obtained by taking the cluster center point as the center of the circle is the zero level set phi of the moving target_k。

The method comprises the following specific steps of obtaining a zero level set of the moving target:

the radius of the zero level set of the moving object is calculated according to the following formula:

R_k＝D_k/N_k

wherein R is_kRadius, D, representing zero level set of moving object_kRepresents the sum of distances from all data points in the k class category in the denoised binary image I to the class cluster center, N_kAnd k represents the serial number of the clustering category obtained by clustering the data points in the binary image I after the noise is removed.

With radius R of zero level set of moving object_kAs a radius, a cluster center s of a kth class obtained by clustering the binary image I after the noise is removed_kMaking a circle as the center of the circle to obtain a zero level set phi of the moving target_k。

In the invention, a k-means clustering algorithm is performed on the binary image without noise to obtain a zero level set of the moving object, and the level set of the moving object is placed in the second frame video image to obtain a schematic diagram as shown in fig. 2 (c). In fig. 2(c), three trolleys are moving targets, and three white circles on the three trolleys are zero level sets of the moving targets. As can be seen from the figure, the zero level set of the moving target is positioned on the moving target and is almost similar to the size of the target, so that the time required for updating the zero level set of the moving target for the subsequent multi-variation level set method is shortened while the moving target can be determined.

Step 6: and updating the zero level set of the moving target by using a multi-variation level set method.

The internal energy of the zero level set of the moving object is calculated according to the following formula:

$W\left(\mathrm{\&phi;}\right)=\frac{\mathrm{\&mu;}}{2}\underset{k=1}{\overset{K}{\&Sigma;}}{\&Integral;}_{\mathrm{\&Omega;}}{\left(\right|\&dtri;{\mathrm{\&phi;}}_k|-1)}^{2}dxdy$

wherein W (phi) represents the internal energy of the zero level set of the moving object, and phi represents the zero level set of the moving object_kμ represents a constraint coefficient for internal energy, the value range of μ is 0 < μ < 0.25, K represents a serial number of a cluster category obtained by clustering data points in the binary image I after noise removal, K is 1,2, …, K represents the total number of the motion targets, ∑ represents summation operation, ^ represents integral operation, and Ω represents the second frame video image I₂The plane set, |, represents the absolute value operation, ▽ represents the gradient operator operation, phi_kZero level set representing moving object, x represents zero level set phi of moving object_kThe coordinate value of the horizontal axis of (a) y represents a zero level set phi of the moving object_kThe ordinate coordinate value of (a).

Obtaining a second frame video image I according to the following formula₂Edge of middle moving object:

$g=\frac{1}{1+|\&dtri;G*I_2{|}^2}$

wherein g represents the second frame video image I₂The edge of the medium moving target, |, represents the operation of taking absolute value, ▽ represents the operation of solving gradient operator, G represents Gaussian kernel, I represents convolution operation₂Representing the second frame of video image of the three consecutive frames of video images.

The length of the zero level set of the moving object is calculated according to the following formula:

$L_g\left(\mathrm{\&phi;}\right)=\underset{k=1}{\overset{K}{\&Sigma;}}{\&Integral;}_{\mathrm{\&Omega;}}g\mathrm{\&delta;}\left({\mathrm{\&phi;}}_k\right)|\&dtri;{\mathrm{\&phi;}}_k|dxdy$

wherein L is_g(phi) denotes the length of the zero level set of the moving object, and g denotes the second frame video image I₂The edge of the middle moving object, phi, represents the zero level set phi of the moving object_kSet of (a) phi_kRepresenting zero level of moving objectSet, K denotes a serial number of a cluster category obtained by clustering data points in the binary image I after the noise is removed, K is 1,2, …, K denotes a total number of moving objects, ∑ denotes a summation operation, ^ denotes an integral operation, and Ω denotes a second frame video image I₂The plane set of (1) represents a Dirac function of a single variable, | · | represents an absolute value taking operation, ▽ represents a gradient operator solving operation, and x represents a zero level set phi of a moving target_kThe coordinate value of the horizontal axis of (a) y represents a zero level set phi of the moving object_kThe ordinate coordinate value of (a).

The area of the moving target region is calculated as follows:

$A_g\left(\mathrm{\&phi;}\right)=\underset{k=1}{\overset{K}{\&Sigma;}}{\&Integral;}_{\mathrm{\&Omega;}}gH(-{\mathrm{\&phi;}}_k)dxdy$

wherein A is_g(phi) denotes the area of the moving object region, and g denotes the second frame video image I₂The edge of the middle moving object, phi, represents the zero level set phi of the moving object_kSet of (a) phi_kRepresents a zero level set of the moving object, K represents a serial number of a cluster category obtained by clustering data points in the binary image I after the noise is removed, K ═ 1,2, …, K represents the total number of the moving object, ∑ represents a summation operation, ^ represents an integral operation, and Ω represents a second frame video image I₂H represents the Heaviside function, x represents the zero level set phi of the moving object_kThe coordinate value of the horizontal axis of (A) and y represents a zero level set of the moving objectφ_kThe ordinate coordinate value of (a).

The external energy of the zero level set of the moving object is calculated according to the following formula:

${\mathrm{\&epsiv;}}_{g,{\mathrm{\&lambda;}}_k,v_k}\left(\mathrm{\&phi;}\right)=\underset{k=1}{\overset{K}{\&Sigma;}}{\mathrm{\&lambda;}}_k{L}_g\left(\mathrm{\&phi;}\right)+\underset{k=1}{\overset{K}{\&Sigma;}}{v}_k{A}_g\left(\mathrm{\&phi;}\right)$

wherein,representing the external energy of the zero level set of the moving object, g representing the second frame video image I₂Edge of medium moving object, λ_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer less than 10, K denotes a serial number of a cluster type obtained by clustering data points in the binary image I after the noise is removed, K is 1,2, …, K denotes a total number of moving objects, v denotes a number of moving objects, and K denotes a number of clusters_kConstraints representing the area of a moving target regionCoefficient phi denotes the zero level set phi of the moving object_kA value of-5 < v_k< 5 and v_kNot equal to 0, ∑ denotes a summation operation, L_g(phi) represents the length of the zero level set of the moving object, A_g(phi) represents the area of the moving object region.

The total energy of the zero level set of the moving object is calculated according to the following formula:

$\mathrm{\&epsiv;}\left(\mathrm{\&phi;}\right)=W\left(\mathrm{\&phi;}\right)+{\mathrm{\&epsiv;}}_{g,{\mathrm{\&lambda;}}_k,v_k}\left(\mathrm{\&phi;}\right)$

wherein (phi) represents the total energy of the zero level set of the moving target, and phi represents the zero level set phi of the moving target_kW (phi) represents the internal energy of the zero level set of moving objects,representing the external energy of the zero level set of the moving object, g representing the second frame video image I₂Contour of medium moving object, λ_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer less than 10, k represents the serial number of the cluster category obtained after clustering the data points in the binary image I after removing the noise, v_kA constraint coefficient representing the area of the moving target region, and the value of the constraint coefficient is-5 < v_k< 5 and v_k≠0。

Carrying out level set evolution on the moving target zero level set according to the following formula to obtain an updated moving target zero level set:

$\frac{\&part;{\mathrm{\&phi;}}_k}{\&part;t}=\mathrm{\&mu;}\&lsqb;{\mathrm{\&Delta;\&phi;}}_k-div\left(\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}\right)\&rsqb;+\mathrm{\&delta;}\left({\mathrm{\&phi;}}_k\right)\&lsqb;\underset{k=1}{\overset{K}{\&Sigma;}}{\mathrm{\&lambda;}}_{k}div\left(g\frac{\&dtri;{\mathrm{\&phi;}}_k}{|\&dtri;{\mathrm{\&phi;}}_k|}\right)\&rsqb;+\underset{k=1}{\overset{K}{\&Sigma;}}{v}_{k}g\mathrm{\&delta;}\left({\mathrm{\&phi;}}_k\right)$

wherein,represents the updated zero level set of the moving object,denotes a derivative operation phi_kRepresents the zero level set of the moving object, phi represents the zero level set of the moving object_kK represents the serial number of a cluster class obtained by clustering data points in the binary image I after noise removal, K is 1,2, …, K represents the total number of moving targets, t represents an iteration step length, mu represents a constraint coefficient for internal energy, the value range of the constraint coefficient is 0 & ltmu & lt 0.25, delta represents a Laplacian operator operation, div (·) represents the divergence of a solved vector, ▽ represents a gradient operator operation, | · | represents an absolute value operation, represents a Dirac function of a single variable, ∑ represents a summation operation, and λ_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer < 10, g denotes a second frame video image I₂Edge of medium moving object, v_kA constraint coefficient representing the area of the target region, the value of which is-5 < v_k< 5 and v_k≠0。

And 7: judging whether the updated moving object zero level set is corresponding to the second frame video image I₂If the edges of the middle moving target coincide, executing the step (8), otherwise, executing the step (6);

and 8: outputting a second frame video image I₂And (5) detecting the moving target.

The effect of the present invention will be further described with reference to simulation experiments.

1. Simulation conditions are as follows:

the simulation of the invention is carried out in a hardware environment with an operating system of windows7, a CPU of Intel (R) core (TM) i5-2400, a basic frequency of 3.20GHz and a memory of 4GB and a software environment of MatlabR2011 b. In the experiment, the iteration step of time is Δ t-5, the constraint coefficient of internal energy is μ -0.2/Δ t, the constraint coefficient of the zero level set length of the moving target is λ -5, and the constraint coefficient of the target area is v-1.5.

2. Simulation content:

the data used in the simulation experiment of the invention are 291,292 and 293 frame video images selected from highway II test video in computer Vision and robotics research, and the size of the video image is 320 x 240 pixels.

3. And (3) simulation result analysis:

the simulation experiment of the invention is to detect the moving target of the video image containing a plurality of moving targets. As shown in fig. 2, fig. 2(a) is a binary image obtained by performing inter-frame difference operation on selected 291,292 th and 293 th frame video images containing a plurality of moving objects in the highway ii test video; fig. 2(b) is a binary image obtained by performing noise removal processing on the binary image of fig. 2(a) by using a morphological function bewareaopen function; fig. 2(c) is a schematic diagram of performing k-means clustering on the noise-removed binary image of fig. 2(b) to obtain a moving target zero-level set, and displaying the moving target zero-level set in a second frame video image to obtain the moving target zero-level set; fig. 2(d) is an effect diagram of moving object detection on a video image containing multiple moving objects, which is obtained by updating the zero level set of the moving objects in fig. 2(c) with multiple diversity level sets.

In fig. 2(d), the cart is the moving target, and the white curve is the zero level set of the moving target. As can be seen from fig. 2(d), the method of the present invention can make the zero level set of the moving target coincide with the edge profile of the moving target, can accurately detect a plurality of moving targets under a complex background, and the obtained detection result has high accuracy.

Claims (4)

1. A multi-target detection method for complex background video images based on a multi-variation level set comprises the following steps:

(1) inputting video images in a video format:

(2) selecting continuous three-frame video image with obvious moving object from input video image, defining it as first frame video image I₁Second frame video image I₂Third frame video image I₃；

(3) Obtaining a binary image:

performing interframe difference on the selected continuous three-frame image by adopting an interframe difference method to obtain a binary image B;

(4) removing interference noise:

denoising other interference noises except the target in the binary image by adopting a morphological function to obtain a denoised binary image I;

(5) obtaining a zero level set of the moving object:

(5a) clustering the binary image I after the noise is removed by adopting a K-means clustering algorithm to respectively obtain the sum of the distances from all data points in a clustering class to the clustering center, the number of all data points in the clustering class and the clustering center;

(5b) taking the result of comparing the sum of the distances from all the data points in the cluster category to the cluster center with the number of all the data points in the cluster category as the radius, and taking the cluster center point as the center of a circle to obtain a circle which is the zero level set phi of the moving target_k；

(6) Updating a moving target zero level set by using a multi-variation level set method:

(6a) the internal energy of the zero level set of the moving object is calculated according to the following formula:

$W\left(\mathrm{\&phi;}\right)=\frac{\mathrm{\&mu;}}{2}\underset{k=1}{\overset{K}{\&Sigma;}}{\&Integral;}_{\mathrm{\&Omega;}}{\left(\right|\&dtri;{\mathrm{\&phi;}}_k|-1)}^{2}dxdy$

wherein W (phi) represents the order of exerciseThe internal energy of the zero level set is marked, phi represents the zero level set phi of the moving object_kμ represents a constraint coefficient for internal energy, the value range of μ is 0 < μ < 0.25, K represents a serial number of a cluster category obtained by clustering data points in the binary image I after noise removal, K is 1,2, …, K represents the total number of the motion targets, Σ represents a summation operation, ^ represents an integration operation, Ω represents a second frame video image I₂The plane set, |, represents the absolute value operation, ▽ represents the gradient operator operation, phi_kZero level set representing moving object, x represents zero level set phi of moving object_kThe coordinate value of the horizontal axis of (a) y represents a zero level set phi of the moving object_kThe coordinate value of the longitudinal axis of (c);

(6b) obtaining a second frame video image I according to the following formula₂Edge of middle moving object:

$g=\frac{1}{1+|\&dtri;G*I_2{|}^2}$

wherein g represents the second frame video image I₂The edge of the medium moving target, |, represents the operation of taking absolute value, ▽ represents the operation of solving gradient operator, G represents Gaussian kernel, I represents convolution operation₂Representing a second frame of video image of the three consecutive frames of video images;

(6c) the length of the zero level set of the moving object is calculated according to the following formula:

$L_g\left(\mathrm{\&phi;}\right)=\underset{k=1}{\overset{K}{\&Sigma;}}{\&Integral;}_{\mathrm{\&Omega;}}g\mathrm{\&delta;}\left({\mathrm{\&phi;}}_k\right)|\&dtri;{\mathrm{\&phi;}}_k|dxdy$

wherein L is_g(phi) denotes the length of the zero level set of the moving object, and g denotes the second frame video image I₂The edge of the middle moving object, phi, represents the zero level set phi of the moving object_kSet of (a) phi_kRepresents a zero level set of the moving object, K represents a serial number of a cluster category obtained by clustering data points in the binary image I after the noise is removed, K ═ 1,2, …, K represents a total number of the moving objects, Σ represents a summation operation, ^ represents an integral operation, and Ω represents a second frame video image I₂The plane set of (1) represents a Dirac function of a single variable, | · | represents an absolute value taking operation, ▽ represents a gradient operator solving operation, and x represents a zero level set phi of a moving target_kThe coordinate value of the horizontal axis of (a) y represents a zero level set phi of the moving object_kThe coordinate value of the longitudinal axis of (c);

(6d) the area of the moving target region is calculated as follows:

$A_g\left(\mathrm{\&phi;}\right)=\underset{k=1}{\overset{K}{\&Sigma;}}{\&Integral;}_{\mathrm{\&Omega;}}gH(-{\mathrm{\&phi;}}_k)dxdy$

wherein A is_g(phi) denotes the area of the moving object region, and g denotes the second frame video image I₂The edge of the middle moving object, phi, represents the zero level set phi of the moving object_kSet of (a) phi_kRepresents a zero level set of the moving object, K represents a serial number of a cluster category obtained by clustering data points in the binary image I after the noise is removed, K ═ 1,2, …, K represents a total number of the moving objects, Σ represents a summation operation, ^ represents an integral operation, and Ω represents a second frame video image I₂H represents the Heaviside function, x represents the zero level set phi of the moving object_kThe coordinate value of the horizontal axis of (a) y represents a zero level set phi of the moving object_kThe coordinate value of the longitudinal axis of (c);

(6e) the external energy of the zero level set of the moving object is calculated according to the following formula:

${\mathrm{\&epsiv;}}_{g,{\mathrm{\&lambda;}}_k,v_k}\left(\mathrm{\&phi;}\right)=\underset{k=1}{\overset{K}{\&Sigma;}}{\mathrm{\&lambda;}}_k{L}_g\left(\mathrm{\&phi;}\right)+\underset{k=1}{\overset{K}{\&Sigma;}}{v}_k{A}_g\left(\mathrm{\&phi;}\right)$

wherein,representing the external energy of the zero level set of the moving object, g representing the second frame video image I₂Edge of medium moving object, λ_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer less than 10, K denotes a serial number of a cluster type obtained by clustering data points in the binary image I after the noise is removed, K is 1,2, …, K denotes a total number of moving objects, v denotes a number of moving objects, and K denotes a number of clusters_kA constraint coefficient for representing the area of the moving target region, phi represents the zero level set phi of the moving target_kA value of-5 < v_k< 5 and v_kNot equal to 0, sigma denotes a summation operation, L_g(phi) represents the length of the zero level set of the moving object, A_g(phi) represents the area of the moving target region;

(6f) the total energy of the zero level set of the moving object is calculated according to the following formula:

$\mathrm{\&epsiv;}\left(\mathrm{\&phi;}\right)=W\left(\mathrm{\&phi;}\right)+{\mathrm{\&epsiv;}}_{g,{\mathrm{\&lambda;}}_k,v_k}\left(\mathrm{\&phi;}\right)$

wherein (phi) represents the total energy of the zero level set of the moving target, and phi represents the zero level set phi of the moving target_kW (phi) represents the internal energy of the zero level set of moving objects,representing the outer part of a zero level set of a moving objectEnergy, g denotes the second frame video image I₂Contour of medium moving object, λ_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer less than 10, k represents the serial number of the cluster category obtained after clustering the data points in the binary image I after removing the noise, v_kA constraint coefficient representing the area of the moving target region, and the value of the constraint coefficient is-5 < v_k< 5 and v_k≠0；

(6g) Carrying out level set evolution on the moving target zero level set according to the following formula to obtain an updated moving target zero level set:

$\frac{\&part;{\mathrm{\&phi;}}_k}{\&part;t}=\mathrm{\&mu;}\&lsqb;{\mathrm{\&Delta;\&phi;}}_k-div\left(\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}\right)\&rsqb;+\mathrm{\&delta;}\left({\mathrm{\&phi;}}_k\right)\&lsqb;\underset{k=1}{\overset{K}{\&Sigma;}}{\mathrm{\&lambda;}}_{k}div\left(g\frac{\&dtri;{\mathrm{\&phi;}}_k}{|\&dtri;{\mathrm{\&phi;}}_k|}\right)\&rsqb;+\underset{k=1}{\overset{K}{\&Sigma;}}{v}_{k}g\mathrm{\&delta;}\left({\mathrm{\&phi;}}_k\right)$

wherein,represents the updated zero level set of the moving object,denotes a derivative operation phi_kRepresents the zero level set of the moving object, phi represents the zero level set of the moving object_kK represents the serial number of a cluster class obtained by clustering data points in the binary image I after noise removal, K is 1,2, …, K represents the total number of moving targets, t represents an iteration step length, mu represents a constraint coefficient for internal energy, the value range of the constraint coefficient is 0 & ltmu & lt 0.25, delta represents a Laplacian operator operation, div (·) represents the divergence of a solved vector, ▽ represents a gradient operator operation, | · | represents an absolute value operation, represents a Dirac function of a single variable, Σ represents a summation operation, and λ represents a sum of a single variable_kA constraint coefficient representing the length of the zero level set of the moving target, and the value of the constraint coefficient is more than 0 and more than lambda_kAn integer < 10, g denotes a second frame video image I₂Edge of medium moving object, v_kA constraint coefficient representing the area of the target region, the value of which is-5 < v_k< 5 and v_k≠0；

(7) Judging whether the updated moving object zero level set is corresponding to the second frame video imageLike I₂If the edges of the middle moving target coincide, executing the step (8), otherwise, executing the step (6);

(8) outputting a second frame video image I₂And (5) detecting the moving target.

2. The multi-target detection method for the complex background video images based on the multi-variation level set as claimed in claim 1, characterized in that: the specific steps of the interframe difference method in the step (3) are as follows:

in the first step, the first frame video image I in the selected continuous three frames video images is used₁Subtracting the second frame video image I₂Obtaining a differential image B¹；

Secondly, using the second frame video image I in the selected continuous three frame video images₂Subtracting the third frame video image I₃Obtaining a differential image B²；

Thirdly, two differential images B¹And B²And (5) carrying out phase AND to obtain a binary image B.

3. The multi-target detection method for the complex background video images based on the multi-variation level set as claimed in claim 1, characterized in that: the K-means clustering algorithm in the step (5a) comprises the following specific steps:

the method comprises the steps of firstly, randomly selecting K central points from a binary image I after noise removal, wherein K represents the total number of moving targets;

secondly, traversing all data points in the binary image I after the noise is removed according to the following formula, and dividing each data point into the nearest central point:

$\min J\left(C\right)=\underset{k=1}{\overset{K}{\&Sigma;}}\underset{x_i\&Element;C_k}{\&Sigma;}{m}_{ki}dist{(x_i,s_k)}^2$

j (C) represents the distance from each data point in the binary image I after the noise is removed to the divided central point, C represents a cluster category set obtained after the data points in the binary image I after the noise is removed are divided, min (·) represents minimum value taking operation, Σ represents summation operation, m (·) represents summation operation_kiA representation coefficient, I represents the serial number of the data points in the binary image I after the noise removal, K represents the serial number of the cluster type obtained by dividing the data points in the binary image I after the noise removal, K is 1,2, …, K represents the total number of moving targets, dist (·) represents a distance function, and x represents the total number of the moving targets_iRepresenting the ith data point, s, in the denoised binary image I_kA cluster center point C representing the kth class obtained by dividing the binary image I after removing the noise_kRepresenting the class category of the kth class obtained by dividing the binary image I after removing the noise, and ∈ representing the operation;

thirdly, calculating the average value of each clustering category obtained after dividing the binary image I after removing the noise according to the following formula, and taking the average value as a new clustering center point:

$s_k=\frac{1}{n_k}\underset{i=1}{\overset{n_k}{\&Sigma;}}{x}_i^k$

wherein s is_kRepresenting the clustering center point of the kth class obtained after the division of the binary image I after the noise removal, k representing the serial number of the class obtained after the division of the data points in the binary image I after the noise removal, n_kRepresents the number of data points in the kth clustering category obtained after the data points in the binary image I after the noise is removed are divided, represents the summation operation,and the data points in the binary image I after the noise is removed are divided to obtain the ith data point in the kth clustering class, and I represents the serial number of the data point in the binary image I after the noise is removed.

4. The multi-target detection method for the complex background video images based on the multi-variation level set as claimed in claim 1, characterized in that: the specific steps of obtaining the zero level set of the moving object in the step (5b) are as follows:

firstly, calculating the radius of a zero level set of a moving target according to the following formula:

R_k＝D_k/N_k

wherein R is_kRadius, D, representing zero level set of moving object_kRepresents the sum of distances from all data points in the k class category in the denoised binary image I to the class cluster center, N_kRepresenting the number of all data points belonging to the kth class in the binary image I after the noise is removed, wherein k represents the serial number of the cluster class obtained after clustering the data points in the binary image I after the noise is removed;

second, using radius R of zero level set of moving object_kRadius, to remove the value of the noiseClustering center s of kth class obtained by image I clustering_kMaking a circle as the center of the circle to obtain a zero level set phi of the moving target_k。